maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

[MachineCopyPropagation] Fix a bug with undef handling when the value is actualy alive. 

Test case will follow.

llvm-svn: 238518
This commit is contained in:
Quentin Colombet 2015-05-28 22:38:40 +00:00
parent cb337035f2
commit 75afbfd4a1
1 changed files with 9 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
llvm/lib/CodeGen/MachineCopyPropagation.cpp
View File

@ -252,11 +252,7 @@ bool MachineCopyPropagation::CopyPropagateBlock(MachineBasicBlock &MBB) {
report_fatal_error("MachineCopyPropagation should be run after"
" register allocation!");
// Treat undef use like defs.
// The backends are allowed to do whatever they want with undef value
// and we cannot be sure this register will not be rewritten to break
// some false dependencies for the hardware for instance.
if (MO.isDef() || MO.isUndef()) {
if (MO.isDef()) {
Defs.push_back(Reg);
continue;
}
@ -270,6 +266,14 @@ bool MachineCopyPropagation::CopyPropagateBlock(MachineBasicBlock &MBB) {
MaybeDeadCopies.remove(CI->second);
}
}
// Treat undef use like defs for copy propagation but not for
// dead copy. We would need to do a liveness check to be sure the copy
// is dead for undef uses.
// The backends are allowed to do whatever they want with undef value
// and we cannot be sure this register will not be rewritten to break
// some false dependencies for the hardware for instance.
if (MO.isUndef())
Defs.push_back(Reg);
}
// The instruction has a register mask operand which means that it clobbers